Abstract: The likelihood ratio computation in Large Vocabulary Continuous Speech Recognition(LVCSR) systems based on continuou density Hidden Markov Model(HMM) is analyzed. The feasibility of using t parallel method to implement the likelihood computation is showed. On basis of this, a fast algorithm for likelihood ratio based on SIMD is proposed, which is implemented by improving likelihood computation modules in HTK3.4 toolkit. Experimental results show this algorithm can speed up the likelihood computation without lowering the accuracy rate of recognition of premise.

Key words: SIMD technology, likelihood computation, Hidden Markov Model(HMM), speech recognition

摘要： 分析基于连续概率密度的隐马尔可夫模型大词汇量连续语音识别系统中的似然率计算方法，阐述运用并行方式实现似然率计算的可行性，并在此基础上，提出一种基于SIMD的似然率快速算法，通过对语音识别工具包HTK 3.4中似然率计算模块的改进实现该算法。实验结果表明，在不降低识别准确率的前提下，该算法能有效加快似然率计算的速度。

关键词: SIMD技术, 似然率计算, 隐马尔可夫模型, 语音识别

CLC Number: 

• TP301.6